On Fri, 2025-06-20 at 15:04 -0400, Antoni Boucher wrote: > Hi. > In rustc_codegen_gcc, we rely on recording::memento::get_debug_string > even though this function should only be used for debugging (because > there are no ways for now to compare rvalues, but this is another, > bigger story) and we faced huge memory usage (+100 GB) because > recording::memento::get_debug_string create memento strings, which > means > they will only be freed at the end of the compilation process. > > Would it be OK to change this so that the responsibility of freeing > the > strings is moved to the caller of gcc_jit_object_get_debug_string? > That might require to remove the caching that is done in > recording::memento::get_debug_string. > > What are your thoughts on this?
I don't like this idea, since it would change the meaning of the API, creating a memory leak in any existing code that uses gcc_jit_object_get_debug_string. I suppose we could add a new entrypoint that creates a client-owned debug string, but it sounds like what you really want is a way to compare rvalues. What sort of comparison do you need on rvalues, and if we provided it, would it solve your memory issues? Hope this is constructive Dave
